Load Shedding CS 240 B notes 1 Load

Load Shedding CS 240 B notes 1

Load Shedding in a DSMS z DSMS: online response on boundless and bursty data streams—How? z By using approximations and synopses and even z Shedding load when arrival rates become impossible z Approximations and Synopses are often used with normal load z Shedding is used for bursty streams and overload situations. 2

Qo. S and Load Schedding z. When input stream rate exceeds system capacity a stream manager can shed load (tuples) z Load shedding affects queries and their answers: drop the tasks and the tuples that will cause least loss z Introducing load shedding in a data stream manager is a challenging problem z Random load shedding or semantic load shedding 3

Problems to Address z. When to shed load y. Overload should be detected quickly z. Where to shed load y. Avoid wasted work y. Upstream Drop Vs. Downstream Drop z. How much to shed y. The magnitude of the drop z. Which tuples to shed 4

Loss-tolerance Qo. S function Loss function is not linear: 5

Value-based Qo. S z. Value-based Qo. S y. Show which values of the output tuple space are most important. y. In a medical application that monitors patient heartbeats y. Extreme values are certainly more interesting than normal ones y. Corresponding higher utility 6

Load Shedding in Aurora z Qo. S for each application as a function relating output to its utility – Delay based, drop based, value based z Techniques for introducing load shedding operators in a plan such that Qo. S isdisrupted the least – Determining when, where and how much load to shed 7

Which Query to drop First? z. Models and algorithms proposed include: y Greedy algorithms, y Fractional Knapsack techniques y. Other OR techniques y Must deal with nonlinearities 8

Load Shedding in STREAM z Formulate load shedding as an optimization problem for multiple sliding window aggregate queries – Minimize inaccuracy in answers subject to output rate matching or exceeding arrival rate z Consider placement of load shedding operators in query plan – Each operator sheds load uniformly with probability pi 9

Window-Oriented Load Shedding Input stream divided into windows of size w z. Use fewer Slides per windows to compute aggregates—tumbles is the extreme case. z Window-based Sampling y Reservoir sampling for incoming tuples y Expiring tuples pose a more difficult problem. 10

Load Shedding by Sampling for Continuous Aggregate Queries on Data Streams: Only random samples are available for computing aggregate queries because of z y Limitations of remote sensors, or transmission lines y Load Shedding policies implemented when overloads occur y When overloads occur (e. g. , due to a burst of arrivals} we can 1. drop queries all together, or 2. sample the input---much preferable z Key objective: Achieve answer accuracy with sparse samples for complex aggregates on windows z Can we improve answer accuracy with minimal overhead? 11

Load Shedding z To cope with bursty arrivals of high-volume data z DSMS has to shed load while minimizing the degradation of the Quality of Service (Qo. S) z The goal then becomes determining: ywhen, where and how much load to shed z An intelligent scheme, can improve the quality of our mining results under bursty arrivals 12

A first Architecture S 1 z Basic Idea: [BDM 04] Query Network y Optimize sampling rates of load shedders for accurate answers. y Find an error bound for each aggregate query. y Determine sampling rates that minimize query inaccuracy within the limits imposed by resource constraints. y This approach works for SUM and COUNT y Generalization to other functions? Sn …. . . ∑ ∑ …. . . Si ∑ Data Stream Load Shedder Query Operator ∑ Aggregate 13

Query Network: arbitrary placement of aggregates and shedder after any aggregate S 1 Sn L 1 L 4 Q 1 Q 4 L 2 Q 2 L 5 Q 3 Q 5 Data Stream Load Shedder Aggregate Operator 14

Generalized Load Shedding in Stream Mill 1. A general framework that achieves optimal load shedding policies, while accommodating: y Different requirements for different users, different query sensitivities, and different penalties. 2. Applicability to a wide spectrum of aggregate functions: y We have formally characterized using a new notion, called reciprocal-error queries. 3. Proposing an extensible architecture that allows UDAs to benefit from the system provided load shedding functions. 4. Significant improvements (in absolute error, false positives, and false negatives) compared to the common uniform approach. 5. We propose an efficient (linear-time) algorithm to handle severe overloads without losing optimality. 15

Goals to Achieve z. Light-weight overhead handling y. React to overload immediately z. Minimizing Qo. S degradation z. Delivering subset results y. Only omitting tuples from the correct answer y. Never produce incorrect answers 16

Prediction & Improvements z A larger class of queries was considered in [LZ 08] y SUM, COUNT, AVG, Quantiles. z Temporal Correlation between answers can be used to improve answer y Example: sensor data y Current answer can be adjusted by the past answers so that: x Low sampling rate current answer less accurate more dependent on history. x High sampling rate current answer more accurate less dependent on history. z A Bayesian quality enhancement module which can achieve this objective automatically and reduce the uncertainty of the approximate answers. 17

Improved Model Using History y the observed answer y the error model P y history of the answer …. . . Query Network z The observed answer à is computed from random samples of the complete stream with sampling rate P. z A bayesian method to obtain the improved answer by combining S 1 Sn ∑ …. . . ∑ ∑ …. . . à History Quality Enhancement Module Si Data Stream Load Shedder Improved answer Query Operator ∑ Aggregate 18

Summary z An error model y. Works for ordered statistics and data mining functions as well as with traditional aggregates, y computationally very efficient y Bayesian quality enhancement method for approximate aggregates in the presence of sampling. y No correction when concept changes are suspected —a two-sample test used to detect suspected changes. 19

References—Sampling and load shedding [Tabul 03] Nesime Tatbul, Ugur Cetintemel, Stanley B. Zdonik, Mitch Cherniack, Michael Stonebraker: Load Shedding in a Data Stream Manager. VLDB 2003, pp. 309 --320. [BDM 04] Brian Babcock, Mayur Datar, Rajeev Motwani: Load Shedding for Aggregation Queries over Data Streams. ICDE 2004: 350 -361. [Tabul 07] Nesime Tatbul, Ugur Cetintemel, Stanley B. Zdonik: Staying FIT: Efficient Load Shedding Techniques for Distributed Stream Processing. VLDB 2007: 159 -170. [LZ 08] Yan-Nei Law and Carlo Zaniolo: Improving the Accuracy of Continuous Aggregates and Mining Queries on Data Streams under Load Shedding. International Journal of Business Intelligence and Data Mining, 2008. [ICDE 2010] Barzan Mozafari and Carlo Zaniolo, Optimal Load Shedding with Aggregates and Mining Queries. In Proceedings of the 26 th International Conference on Data Engineering (ICDE 2010), Long Beach, California, USA, March 1 -6, 2010. 20